Smoking Withdrawal Dynamics in Unaided Quitters

Thomas M. Piasecki University of Wisconsin Medical School and University of Wisconsin---Madison

Raymond Niaura, William G. Shadel, Michael C. Fiore and Timothy B. Baker David Abrams, and Michael Goldstein University of Wisconsin Medical School and Brown University University of Wisconsin---Madison

Considerable research shows that withdrawal severity is inconsistently related to smoking cessation outcomes. This may result from measurement problems or failure to scrutinize important dimensions of the withdrawal experience. Two recent studies demonstrated that withdrawal elevation and variations in the time course of withdrawal were related to relapse in smokers treated with the nicotine patch (T. M. Piasecki, M. C. Fiore, & T. B. Baker, 1998). This article reports a conceptual replication and extension of those findings in unaided quitters. Evidence for temporal heterogeneity was found across different types of withdrawal symptoms. Patterns or slopes of affect and urge reports over time predicted smoking status at follow-up, as did mean elevation in withdrawal symptoms. These results suggest that affect and urge withdrawal symptoms make independent contributions to relapse and that relapse is related to both symptom severity and trajectory.

The study of withdrawal is at a watershed. Traditional models of because measures of withdrawal have not covaried consistently drug dependence and drug motivation have accorded withdrawal a with relapse vulnerability and other purported indices of depen- central role in motivating addictive drug use and relapse (e.g., dence. Theoreticians have offered a litany of observations that Isbell & White, 1953; Seevers, 1962; Wikler, 1973). In addition, challenge withdrawal-based accounts. These include the fact that for many years, the presence of withdrawal symptoms has consti- there is an inconsistent relation between the tendency of a drug to tuted a central criterion for the diagnosis of a drug dependence support habitual use and the severity of the drug's withdrawal disorder (American Psychiatric Association, 1994). However, re- syndrome (e.g., Jaffe, 1980). Also, drug-dependent persons fre- cent models of dependence and drug motivation have almost quently do not identify withdrawal as an instigator of relapse universally de-emphasized withdrawal as a central explanatory (Marlatt, 1985), and relapse often occurs long after the cessation of construct (cf. Heyman, 1996; Marlatt, 1985; Robinson & Berridge, drug use, when withdrawal symptoms should have abated (Bran- 1993; Vuchinich & Tucker, 1988). This has transpired largely don, Tiffany, Obremski, & Baker, 1990). Finally, the extent to which treatments suppress withdrawal does not correspond well with their ability to produce abstinence (Jorenby et al., 1995; Thomas M. Piasecki and Timothy B. Baker, Center for Tobacco Re- Transdermal Nicotine Study Group, 1991). search and Intervention, University of Wisconsin Medical School, and Various factors might account for the weak relations between Department of Psychology, University of Wisconsin--Madison; Raymond withdrawal and relapse. For instance, it may be the case that Niaura, William G. Shadel, David Abrams, and Michael Goldstein, Divi- sion of Behavioral and Preventive Medicine, The Miriam Hospital, Brown smoking withdrawal symptoms have little impact on cessation and University; Michael C. Fiore, Center for Tobacco Research and Interven- the maintenance of abstinence. Alternatively, inadequate concep- tion and Department of Medicine, University of Wisconsin Medical tualization and assessment strategies may have caused researchers School. to overlook important dimensions of the withdrawal phenomenon This research was supported in part by National Heart, Lung and Blood (Gilbert et al., 1998; Piasecki, Kenford, Smith, Fiore, & Baker, Institute Grant HL32318 and by National Institute on Drug Abuse Grant 1997). Prior to abandoning withdrawal as a vital explanatory DA07580-03. We gratefully acknowledge Claudia Morelli, Suzanne Sales, construct, it seems prudent to pursue novel approaches to the study Adrienne McParlin, Cheryl Eaton, and Alicia Fontes for their help in data collection. of withdrawal that may yield additional information about its Correspondence concerning this article should be addressed to Thomas correlates and consequences (Meehl, 1978; Patten & Martin, M. Piasecki, Center for Tobacco Research and Intervention, 7275 Medical 1996a). Indeed, investigators have begun to adopt new and inno- Sciences Center, 1300 University Avenue, Madison, Wisconsin 53706, or vative approaches to the study of withdrawal (e.g., Epping-Jordan, Raymond Niaura, The Miriam Hospital, Division of Behavioral and Watkins, Koob, & Markou, 1998; Gilbert et al., 1998; Keenan, Preventive Medicine, 164 Summit Avenue, Providence, Rhode Island Hatsukami, Pentel, Thompson, & Grillo, 1994; Shiffman et al., 02906. Electronic mail be sent to [email protected] or 1997; Swan, Ward, & Jack, 1996). Raymond Niaura @brown.edu.

74 SMOKING WITHDRAWAL 75

When a mismatch between strong theory and empirical findings nections between symptom trajectories and an array of affect and exists, measurement issues are a natural object of scrutiny (Serlin dependence indices, and we use a variety of strategies to begin & Lapsley, 1985); investigators should be reluctant to discard teasing apart the connections between postcessation smoking and venerable theories if there is a possibility that empirical discrep- withdrawal patterns. We assess the clinical relevance of atypical ancies might stem from inadequate measures. Hence, it is not withdrawal by examining the connections between categorical surprising that smoking researchers have begun to reexamine (cluster based) and continuous (simple slope) measures of with- withdrawal instruments (Patten & Martin, 1996b; Welsch et al., drawal trajectory and relapse. Central theoretical principles guid- 1999). Another tactic that can be used to reevaluate withdrawal is ing our research on withdrawal heterogeneity posit that the trajec- to start with withdrawal measures that are presumed valid and then tory of withdrawal symptoms reflects the responsivity of affective analyze them in novel ways. processing systems to pharmacologic and nonphannacologic stim- We have recently reported results from two studies examining ulation (Piasecki et al., 1997, 1998) and that affects provide a variability in the time course of withdrawal symptoms that reflect readout of motivated behavior (e.g., Baker, Morse, & Sherman, this approach to measurement evaluation (Piasecki, Fiore, & 1987; Buck, 1993). These principles give rise to the prediction that Baker, 1998). In these studies, we clustered quitters according to the affective components of the withdrawal syndrome should show the shapes of their withdrawal profiles and demonstrated that the strongest relations with relapse. atypical withdrawal, characterized by late peaks or unremitting symptoms, was quite common. Such withdrawal profiles are Method termed atypical because they diverge from the prototype generated by traditional models of withdrawal (Hughes, Higgins, & Hat- Participants and Procedure sukami, 1990). Atypical profiles were common, occurring in smokers who had lapsed and those who were abstinent and occur- Data for this report were drawn from a study of unaided quitters conducted at The Miriam Hospital in Providence, Rhode Island (Shadel et ring in smokers who received diverse interventions. Data were al., 1998). To be eligible for the study, participants had to: (a) be be- drawn from two clinical trials of the nicotine patch, and the tween 18 and 75 years of age; (b) have no decrease in the number of withdrawal scale and assessment schedules used in these trials cigarettes or nicotine content of their usual brand of cigarettes by more than conformed to common research practice. half in the previous 3 months; (c) have no use of tobacco products other Atypical profiles were not only shown to be fairly common, but than cigarettes; (d) report motivation to quit as at least 5 on a 10-point they were also related to relapse; participants with atypical profiles scale; (e) not be enrolled in any other smoking cessation program or use were more likely to relapse than were other smokers. This relation any other quitting methods (e.g., patch, gum); (f) not attempt, if female, to could involve a variety of mechanisms. For instance, smokers with become pregnant; (g) not be currently treated for ongoing psychiatric worsening symptoms might be more likely to return to smoking in disorders; and (h) not be taking medication that might interfere with order to manage their worsening symptoms. Smokers who lapse psychophysiological assessments (e.g., beta blockers). Participants visited the laboratory a total of eight times during the study. might experience an exacerbation in withdrawal due to the lapse, Three sessions occurred prior to the quit date. At the first session, partic- and this might motivate further smoking and relapse. Thus, al- ipants completed a variety of self-report measures and gave blood samples. though little is known about the mechanisms that link withdrawal At a second session, held the next day, participants underwent various profiles to relapse, there is evidence that profile shape carries experimental manipulations of smoking motivation (see Shadel et al., motivational meaning. 1998). The third session, scheduled 3 days after the experimental session, Although preliminary research suggests that smoking with- was a brief check-in meeting where participants completed a withdrawal drawal heterogeneity is real and consequential, vital questions assessment and were given a copies of the American Lung Association's remain about the phenomenon and its generality requires ex- "Freedom from Smoking" self-help pamphlet. The quit day for all partic- ploration. For instance, our previous research was conducted ipants was scheduled for 4 days after the check-in meeting. Participants with participants who were undergoing intensive formal cessa- reported to the lab for brief assessments (including withdrawal assess- tion treatments. Because smokers who volunteer for such treatments) on the quit date and on Days 2, 7, 14, and 30 postquit. Other than the self-help pamphlet, participants were offered no formal behavioral ments tend to differ from other smokers (Fiore et al., 1990; support or pharmacological treatment for smoking cessation. Participants Lichtenstein & Hollis, 1992), it is important to establish the received $170 for completing the study. generalizability of the results with unaided quitters. In addition, The original sample included 183 individuals, however, the profile it is vital to determine whether withdrawal profile type is methods used in this research required that complete withdrawal data be significantly related to relapse when participants' postcessation available for all participants at each follow-up time point. This criterion smoking/lapse status is statistically controlled. If withdrawal eliminated 24 individuals, leaving 159 cases for analysis. The analyzed profile is merely a proxy for lapse status, then profile informa- sample was 56% female and 95% Caucasian with a mean age of 42.9 years tion should contribute nothing to statistical models of relapse (SD = 12.1). The analyzed sample reported a mean Fagerstrrm Tolerance once lapse status is taken into account. Finally, the withdrawal Questionnaire (FrQ; Fagerstrrm, 1978) score of 6.5 (SD = 1.8), a history measure used in the previous research was a brief eight-item of 3.3 serious quit attempts (SD = 2.3), and a typical smoking rate of 24.2 (SD = 9.5) cigarettes per day. A series of analyses comparing analyzed scale laden with items tapping affect. A broader measure with participants with excluded sample members did not reveal significant multiple-item scales tapping distinct symptom domains may differences between the groups on any demographic or dependence reveal different results across the different domains. variables. In this article, we examine symptom trajectories in a variety of The sample analyzed was not limited to only those individuals who withdrawal subdomains in a sample of unaided quitters. To char- maintained continuous abstinence throughout the follow-up period. acterize temporal variance in withdrawal reports, we explore con- Eliminating participants who slip or lapse may restrict the range of both 76 PIASECKI ET AL. withdrawal and relapse data, attenuating relations between them (e.g., relates, examined an array of affective, demographic, and dependence Gilbert et al., 1998). Instead, we opted to conduct analyses to assess the variables to see whether any of these foreshadowed postcessation impact of intratrial smoking on our findings (see following discussion). symptom patterning.

Measures Descriptive Analyses Factor score evaluation. The withdrawal measure used in this re- Four types of measures were used in this research: measures of with- search contained a mixture of items from different scales, and item drawal, postcessation cigarette counts, end-of-study abstinence, and base- scores were transformed to yield nonredundant symptom measures. line smoker characteristics. Before forming subgroups of smokers based upon these factor score scales, we sought to determine whether the measures showed aggregate Withdrawal temporal patterns consistent with those shown in the descriptive liter- ature. Mean profiles computed across the entire sample for each factor The primary measure of withdrawal symptomatology in this research score scale were plotted. In essence, these plots assessed whether the included items from the Hughes and Hatsukami (1986) and Shiffman and variance unique to each symptom domain met a minimum validity Jarvik (1976) withdrawal scales, along with additional items tapping urge/ standard as a measure of withdrawal. craving. An exploratory factor analysis with principal components extrac- Profile clustering. The basic profile methodology from our previous tion and varimax rotation performed on the quit-day data suggested that studies was used with these data. In brief, participants' withdrawal three factors influenced the manifest item intercorrelations. The first factor ratings for Days 1, 2, 7, 14, and 30 were considered to constitute their largely comprised items related to negative affect (Affect), the second dynamic profile. Profiles were standardized casewise to remove eleva- factor largely concerned urges/cravings (Urge), and the third factor in- tion and scatter, leaving only shape information (i.e., the patterns of ups cluded items concerning sleep disturbance/energy (Sleep/Energy). The and downs; Cronbach & Gleser, 1953). Then, the 159 standardized withdrawal items and their loadings on the factors to which they were profiles were submitted to a hierarchical agglomerative cluster analysis assigned are given in the Appendix. These symptom subdomains were using Ward's (1963) minimum variance procedure and the squared considered separately, and dynamic patterns in each were examined using Euclidian distance measure. In this procedure, each resulting cluster cluster analysis. To create symptom subscales that were conceptually pure represents a group of individuals with similar temporal trajectories on and distinct, the factor score coefficient matrix derived from the quit-day the measure under consideration (e.g., Prochaska, Velicer, DiClimente, data factor analysis was used to compute quasi-orthogonal subscales at all Gaudagnoli, & Rossi, 1991). This procedure was performed three times, time poiats. once for profiles from each of the symptom domains. Thus, the entire analyzed sample was partitioned three separate times, and each partic- Postcessation Cigarette Counts ipant was assigned to an affect scale cluster, an urge scale cluster, and a sleep/energy scale cluster. At each postcessation study visit, participants were asked to recall Stopping rules for hierarchical cluster analysis remain the focus of retrospectively how many cigarettes they had smoked each day since the considerable research and debate (e.g., Krolak-Schwerdt & Eckes, last study visit. 1992; Overall & Magee, 1992; Schweizer, 1992), and no single quan- titative or heuristic algorithm has clearly emerged as superior. Absent Abstinence an empirically validated criterion, we opted to use a simple, practical heuristic; we retained the maximum number of clusters with no fewer Participants were considered to be abstinent at the end of the study if than 20 participants in any cluster (Morral, Iguchi, Belding, & Lamb, they reported smoking zero cigarettes in the 7 days preceding the Day 30 1997), in conjunction with visual inspection of the dendrogram and assessment and if they provided a saliva sample containing less than 20 fusion coefficients. Such a method allowed for participant differentia- ng/ml cotinine. tion while restricting the retention of trivially small groups. This rule is somewhat arbitrary, but its validity is better evaluated by examination Baseline Smoker Characteristics of its usefulness for organizing other aspects of the data than by its mathematical rigor or sophistication (Morral et al., 1997; Skinner, At the first precessation visit, participants completed a series of paper- 1981). and-pencil measures that tapped smoking history/nicotine dependence and Prototypic-atypical classification. For analyses in subsequent phases mood. Smoking history/dependence measures included the FTQ, age of of this study, members of atypical clusters were combined. This was smoking initiation, regular smoking rate, number of serious quit attempts, done for both theoretical and statistical reasons. First, whether or not an and plasma nicotine and cotinine. Mood measures were the Positive and individual displays an atypical withdrawal trajectory may be more Negative Affect Schedule (PANAS; Watson, Clark, & Tellegen, 1988), the important than the specific form of that trajectory (Piasecki et al., Center for Epidemiological Studies Depression Scale (CESD; Radloff, 1997). Theory and prior research (Piasecki et al., 1998) suggest that if 1977), and the Perceived Stress Scale (PS; Cohen, Kamarck, & Mermel- stein, 1983). A variety of other measures were collected but were not used for the purposes of this article. 1 These factor scores were, of course, only quasi-orthogonal, because a factor score coefficient matrix derived at one time point was used to adjust Statistical Analyses scores at all time points. These coefficients do not yield truly orthogonal subscales at all time points, because they capitalize on indiosyncracies in A three-phase analytic strategy was adopted for this article. The the quit day data. Nonetheless, it seemed most reasonable to apply a descriptive phase served to characterize temporal variability in symp- consistent mathematical rule for adjusting item scores at all time points. On tom subdomains and to form participant subgroups that would become the quit day, the Affect and Sleep/Energy subscales were intercorrelated at the bases for subsequent analyses. The second analytic phase involved .58, whereas their factor scores were correlated at -.09. Comparable screening groupings from the first phase for clinical relevance by using diminution in subdomain intercorrelations was observed across subdomain a series of logistic regression analyses. The final phase, external cor- pairs and time points. SMOKING WITHDRAWAL 77 smokers have withdrawal symptoms that either increase or remain Contribution of nonwithdrawal variables. After the multivariate with- elevated, they are at higher risk of relapse than smokers who have drawal model was built, we sought to determine whether any of the smoker withdrawal symptoms that steadily abate. Collapsing across atypical characteristics measured at baseline would improve the model. In separate profile types also increases statistical power. Although we collapse logistic regression analyses, each of the smoker characteristic variables was participants with atypical profiles for theoretically motivated statistical added to the best-fitting withdrawal model at the last step. All continuous tests, we graphically depict symptom profiles on a group-by-group gasis predictors were tested for linearity of the logit; where the assumption of to convey maximal descriptive information. linearity appeared to be violated, continuous variables were converted to Cluster analyses generated a family of profiles for each factor score polytomous form and were tested again. The model chi-square improve- scale that was analyzed. To designate a prototypic profile (with a ment statistic was examined in each analysis to determine whether any steadily descending course) from each family of profiles in an objective nonwithdrawal variable significantly augmented the withdrawal model. manner, we first identified mean withdrawal ratings for the relevant Influence of postcessation cigarette use. We conducted several analy- time points from the prototypical group in a previous study (i.e., Study ses to examine the extent to which the differing withdrawal patterns and 1; Piasecki et al., 1998). S correlations with this historical prototype their associations with relapse were attributable to participants' smoking were then computed for all group profiles that were derived from the during the follow-up period. The first strategy was to perform a series of present data. For each scale type, the group with the highest S corre- repeated measures analyses of variance (ANOVAs) using cigarette counts lation (i.e., a product-moment correlation computed between group as the dependent measure, comparing typical and atypical cluster members profiles across time points; Cattell, 1966) was considered the prototype. on the pattern and amount of cigarettes smoked during the trial. Only Dummy coding was used to contrast prototypic and atypical clusters in participants who reported smoking at least one cigarette during the cessa- all subsequent analyses. tion period were included in these analyses. Separate analyses were performed for each cluster solution. Assessing Clinical Relevance Second, logistic regression models were built to assess whether any withdrawal parameter could improve prediction of relapse after a variable A series of logistic regression analyses were performed using Day 30 representing lapse occurrence (i.e., the participant reported smoking at least point-prevalence relapse as the dependent measure. These were con- one cigarette from Day 1-21 of the quit attempt) during the trial was forced ducted to ascertain (a) withdrawal parameters that were associated with into the model at a prior step. Model building followed the same strategy relapse; (b) the best combination of withdrawal parameters for model- that was used to construct the multivariate withdrawal model. At Step 1, ing relapse; (c) the incremental improvement made to the final with- lapse status was entered; at Step 2, all withdrawal parameters retained in drawal model, if any, by nonwithdrawal variables; and (d) the influence of postcessation smoking on relations between withdrawal parameters the f'mal withdrawal model were entered as a block to form a saturated and relapse. Multivariate logistic models were constructed following model. Withdrawal variables that appeared to be weak predictors were then the strategy advocated by Hosmer and Lemeshow (1989). Four with- individually deleted, unless deletion worsened overall model fit. Because drawal parameters were considered as predictors in these analyses: the postcessation slips are strongly and consistently related to relapse at distant three dummy variables, representing the typical-atypical distinction follow-up points (e.g., Gourlay, Forbes, Marriner, Pethica, & McNeil, derived from each cluster analysis, and an elevation variable, repre- 1994; Kenford et al., 1994), this represents a fairly conservative test of the senting each participants' mean level of reported global distress (total hypothesis that withdrawal patterns contain unique motivational scale score) across the assessment period. The elevation variable was information. included based on a priori rationales. First, it was a strong predictor of A final, rather more descriptive approach to this problem involved outcome in our previous studies (Piasecki et al., 1998). Second, it splitting each cluster into subgroups based upon latency to first postces- controls for the impact of symptom severity per se, which allowed us to sation cigarette puff. As many as six groups could be formed for any isolate the predictive validity of profile shape. cluster. One group represented individuals who reported continuous absti- Screening candidate predictors. Before building logistic models, we nence across the 30 days, and five groups represented various postcessation assessed the overlap of the prototypic-atypical groupings for each symp- assessment visits at which the first report of smoking was given (first lapse tom measure using chi-square tests, because redundant classifications reported at Day 1, Day 2, Day 7, Day 14, or Day 30). Once groups were would influence the interpretability of multivariate models. Individual formed, symptom profiles were plotted for each group to depict the degree withdrawal variables were then screened in univariate logistic models of symptom pattern coherence within each cluster. This coherence was also before multivariate modeling. The main effects of each withdrawal param- quantified using coefficient alpha (treating subgroups as items and occa- eter were assessed via separate logistic regression models. In each model, sions as subjects) to provide a summary index of the extent to which the dependent variable was Day 30 point-prevalence relapse rates, as symptoms jointly wax and wane across groups within a cluster. If post- defined earlier, and a withdrawal parameter was entered as the sole cessation smoking events precipitate atypical withdrawal, then we would predictor. expect to see that groups differing in the latency to the first lapse show Multivariate logistic model Withdrawal variables displaying at least systematically different patterns of withdrawal (e.g., time-shifted peaks), modest relations with the outcome measure (p < .25) were considered even though they happen to be in the same cluster, and that only a small for inclusion in a multivariate withdrawal model. Once the pool of family of lapse histories are represented in each cluster (for example, just candidate predictors was determined, a multivariate model containing participants lapsing early in the quit attempt). all of the retained variables (i.e., a saturated model) was fitted. The Symptom slopes and relapse. We use cluster analyses in this research Wald statistic and other estimated coefficients were examined in the because it systematically replicates our prior findings and because cluster- saturated model, and variables that did not contribute to the model were ing is a descriptively rich method of data reduction. Nonetheless, cluster deleted. After a variable was deleted, the likelihood ratio statistic was analysis involves subjective judgments that raise concerns about the reli- computed to determine whether the deletion resulted in a decrement in ability of the procedure. In addition, clustering masks intracluster hetero- overall model goodness of fit; if this occurred, the predictor was geneity. We have argued in previous work that cluster membership predicts retained. This process was performed iteratively until a final model, smoking outcomes because the clusters are formed on the basis of symp- from which no other variables could be deleted, was obtained. The tom trajectory and this has motivational meaning. This suggests that the use overall goodness of fit of this final model was then assessed via the of simple symptom slopes across the postcessation period should replicate Hosmer-Lemeshow ~ statistic. some of the cluster solution findings. Strong negative slopes would indicate 78 PIASECKI ET AL. symptomatic relief and should be associated with abstinence, whereas nicotine dependence measures, which were broadly conceived (Skinner, near-zero or positive slopes would suggest arrested or exacerbating with- 1990). Variables considered in this set of comparisons were: plasma drawal and should be associated with relapse. Finding a relation between nicotine, plasma cotinine, FTQ, age of smoking initiation, regular smoking slopes and outcomes would suggest that symptom-relapse connections are rate, and number of serious quit attempts. The second set of comparisons not dependent upon a fortuitous cluster solution. focu~gxl on affect-relevant baseline characteristics. Variables included in To test this idea, we computed rough slope indices by taking a simple this set were both subscales of the PANAS, the CESD, and the PS. difference score for each participant, subtracting postcessation Day 2 Associations between baseline measures and symptom slopes and elevation scores in each symptom domain from Day 30 scores. These slope measures were tested using zero-order correlations, whereas while associations with were then substituted for the prototypic-atypical group distinction in the prototypic-atypical cluster distinctions were assessed using t tests. three logistic regression models (multivariate withdrawal model, withdrawal model + nonwithdrawal predictors, lapse-adjusted model; deriva- tion of these models was described earlier) to determine whether symptom Results trajectory remained associated with relapse when continuously scaled representations were used. Slope measures and cluster distinctions clearly reflect somewhat differ- Descriptive Analyses ent information about the symptom trajectory. For instance, slope measures would be expected to be relatively insensitive to transient increases in Factor Score Evaluation symptoms occurring in the middle of the cessation period, but this information would influence clustering. A final set of logistic models assessed The time course, averaged across the entire sample, for each whether prototypic-atypical cluster distinctions contributed to prediction transformed symptom subscale, is depicted in Figure 1. In the plot, of relapse after elevation and slope measures were forced into the model at profiles are transformed into a standard z-score space to emphasize an earlier step. These analyses were performed only for symptom classes their shape and to foster comparability. The Affect and Urge that were related to relapse, and they address whether there is any unique value to clustering. subscales generally conformed to the the expected transient temporal pattern expected of classic withdrawal symtoms (Hughes & Hatsukami, 1986; Hughes et al., 1990), although Affect scores had External Correlates not resolved to baseline levels by the end of follow-up. However, A series of analyses tested whether symptom variables that were related the unique variance in the Sleep/Energy items did not show the to clinical outcome in the multivariate withdrawal models were related to classic temporal signature of a withdrawal measure. Scores on this baseline smokers' characteristics. One set of comparisons focused on scale were highest at baseline and were lower than baseline

Affect -...Z~... Sleep/Energy -D-" urge Z~ • I-1.,, '•. "~ '.. "~ LU O [] / / / / / I-1" A ...... ::~ ~ ...... A 441 m 4 • LJxl--I" ~. ' .... ~' ...... -1 //0 ZX'"'" ...... [] ...... []

-2 I I t I I I I B1 B2 QD 2 7 14 30

SESSION

Figure 1. Aggregate time course of the individual symptom measures used in this article. Data points represent mean ratings from the whole sample for each factor score at the corresponding time point. Profiles are depicted in z-score space to eliminate differences in scaling. B1 = 8 days precessation; B2 = 4 days precessation; QD = quit day. Remaining axis labels represent the number of days postquit. SMOKING WITHDRAWAL 79 throughout the abstinence attempt. 2 Despite their unusual aggre- The elevation variable, model h,2(1, N = 159) = 12.55, p < gate pattern, we chose to include Sleep/Energy factor scores in .001, affect classification, model X2(1, N = 159) = 5.82, p = .01, subsequent analyses to characterize temporal variability in this and urge classification, model )(2(1, N = 159) = 5.48, p = .01, measure and to assess its clinical relevance. were all significantly associated with relapse in univariate logistic models. Sleep/energy classification was not associated with relapse, model )(2(1, N = 159) = 0.06, ns. Profile Clustering Seventeen (34%) of the participants assigned to the prototypic affect group were counted as abstinent at the end of the trial, The three panels of Figure 2 display in standardized form the compared to 18 (17%) of atypical affect group members, X2(1, group profiles derived by submitting each withdrawal subscale N = 159) = 6.11, p = .01. Thirty-one (26%) of the participants measure to cluster analysis. 3 In all panels of Figure 2, the proto- assigned to the prototypic urge group were counted as abstinent, typical cluster is Cluster I and is depicted as a solid line. Four compared to 4 (10%) of atypical urge group members, X2(1, N = clusters emerged for Affect factor scores and Sleep/Energy factor 159) = 4.84, p < .05. Relapse rates for sleep/energy typical and scores, whereas two clusters were retained for Urge factor scores. atypical groups were very similar; they were 10 (21%) and 25 Plotting the cluster means against time revealed substantial heter- (22%), respectively, X2(1, N = 159) = 0.56, ns. ogeneity in patterns of withdrawal symptomatology. For all measures, notable proportions (25%-70%) of the sample reported withdrawal trajectories that differed markedly from the prototyp- Multivariate Withdrawal Model ical pattern seen in group-level studies. Affect factor scores. The prototypical group (Cluster I) con- Based on the univariate findings, all of the withdrawal param- tained 50 (31%) participants and showed a rise in symptoms from eters except the sleep/energy classification were considered as the quit day to Day 2 and then a sharp diminution in symptom- candidate predictors in the multivariate model. Elevation, affect atology thereafter. Clusters H-IV comprised 36 (23%), 42 (26%), classification, and urge classification were all significant predic- and 31 (19%) participants, respectively. As is seen in Figure 2, tors when entered simultaneously in a saturated logistic model, so Cluster I showed the most consistent improvement in symptoms no variable was deleted. The top portion of Table 1 summarizes the after Day 2. estimated logistic regression coefficients associated with this Urge factor scores. Cluster I contained 118 (74%) partici- model. The odds ratios reported in Table 1 indicate that, when pants, and Cluster H contained 41 (26%) participants. Again, overall symptomatology and urge classification are controlled, Cluster I members were distinctive in that their Urge scores atypical affect participants had over 4.2 times the odds of being declined steadily, dropping below baseline Urge reports by Day 7 counted as relapsed by Day 30. Likewise, atypical urge members of the quit attempt. had 3.9 times the odds of relapse. A 1-SD increase in overall Sleep/Energy factor scores. The prototype, Cluster I, in- symptom elevation was associated with a 2.7-fold increase in the cluded 48 (30%) participants, whereas Clusters II-IV contained 39 odds of relapse. The Hosmer-Lemeshow fit statistic for this final (25%), 44 (28%), and 28 (18%) participants, respectively. Figure 2 model suggests that it fit the data adequately, ~(8, N = shows that only Cluster I had the characteristic transient pattern, 159) = 11.87, p = .16. with scores increasing slightly above baseline on the quit date and dropping below baseline levels at all subsequent assessments. 2 The temporal pattern of the transformed sleep/energy scores might be interpreted as a simple offset effect (Hughes et al., 1990). However, inspection of the raw Sleep/Energy scale scores (a simple sum of the Assessing Clinical Relevance relevant item scores) did show transient increases in these symptoms after the quit date, a finding that is consistent with other research examining raw Screening Candidate Predictors scores on these symptoms (e.g., Jorenby etal., 1996). We suspect that this discrepancy results from the specification of orthogonal factors. In other The affect classification was not redundant with groupings words, raw scores on the items dominating the Sleep/Energy factor may based on the remaining symptoms. Thirty-seven (74%) of the reflect general withdrawal malaise, but this effect is mitigated when factor prototypic affect smokers were also assigned to the prototypic urge scores are stripped of variance shared with affect and urge. To some extent, group, whereas 28 (26%) of the atypical affect smokers were this finding may reflect component overextraction (Wood, Tataryn, & counted in the atypical urge group, )(2(1, N = 159) = 0.002, ns. Gorsuch, 1996). However, there is some evidence that objective measures Seventeen (34%) of the prototypic affect smokers were counted as of sleep quality poorly correspond with self-report assessments (Wetter, prototypic in the sleep/energy analysis, and 78 (72%) of atypical Fiore, Baker, & Young, 1995), a finding consistent with the interpretation affect smokers were counted as atypical, X2(I, N = 159) = 0.50, that raw scores on sleep/energy items may be driven to some degree by method variance and general discomfort. ns. However, there was some overlap between the urge and sleep/ energy classifications. Twenty-nine (25%) of the prototypic urge 3 The goal of our cluster analyses was to subdivide the total sample into smaller groups that were comparatively homogeneous with respect to the smokers were counted as prototypic in the sleep/energy analysis, shape of their withdrawal profiles. To assess whether this was achieved, we whereas 22 (54%) of atypical urge smokers were classified as examined the variability in withdrawal ratings before and after clustering. atypical, X2(1, N = 159) = 6.84, p < .01. Despite this overlap, we Across cluster solutions and time points, the standard deviations of stan- decided to retain both groupings as potential independent variables dardized symptom measures were generally reduced between 20%-40% for model building, because 108 (69%) of the total sample fell into after clustering, suggesting that the analyses succeeded in grouping to- off-diagonal cells of the contingency table. gether individuals with similar withdrawal trajectories. 80 PIASECKIET AL.

3 m

2 - "* ~ %-" D.~ ,,t s 1 - / 0 go Z ".. 7\ ",,' 0 - ...... : ~.._.~ ; "'.~

-1 --

-2 I I I I I I I B1 B2 QD 2 7 14 30 3-

2- i 1 - O I ...... l~, ..... II ..... [] .... III i 0- .... ~- .... IV

-1 --

-2 I I I I I I I B1 B2 QD 2 7 14 30 3 -

0 0 2 -

>. 1 - • "'". I "/ %"x /~3" .... "V i 0 -

~" . • .. . ..~.

f,,,,I -1 --

-2 -- I I I I I I I B1 B2 QD 2 7 14 30 SESSION

Figure 2. Standardized group symptom profiles derived from cluster analyses. For each measure, the prototypic group is designated Cluster I and is depicted as a solid line with open circles. B1 = 8 days precessation; B2 = 4 days precessafion; QD = quit day. Remaining axis labels represent the number of days postquit.

Contribution of Nonwithdrawal Variables associated with a 1.7-fold increase in the odds of relapse. With plasma nicotine in the model, the odds ratios for the classification Of all baseline smoker characteristics (plasma nicotine, plasma variables were altered slightly, but all withdrawal parameters re- cotinine, b'TQ, age of smoking initiation, regular smoking rate, mained significantly associated with the abstinence criterion. The number of quit attempts, baseline PANAS, CESD, and PS), only Hosmer-Lemeshow fit statistic for this final model suggests that plasma nicotine, model )(2(1, N = 157) = 4.98, p < .05, and the model fit the data well, ~(8, N = 157) = 6.62, p = .58. plasma cotinine, model X2(I, N = 157) = 4.39, p < .05, significantly improved the model when they were added separately. Influence of Postcessation Cigarette Use These measures were significantly correlated with one another (r = .67, p < .001). Therefore, in building the combined model, Repeated measures ANOVAs that compared typical and atypi- we opted to include only plasma nicotine, the slightly stronger cal clusters on the pattern and amount of cigarette use across the predictor. The middle portion of Table 1 displays the results of this cessation period did not reveal significant main effects or Clus- logistic model. A 1-SD increase in plasma nicotine at baseline was ter x Time interactions for any subscale grouping. SMOKING WITHDRAWAL 81

Table 1 cluster subgroups suggests that the cluster analyses performed as Logistic Regression Results Assessing Relations Between they are purported to, that is, they isolated homogeneous sub- Withdrawal Variables and Relapse groups of smokers who showed the same symptom shape regardless of when they started smoking. Second, the fact that smokers Predictor [3 SE (13) Wald p OR with diverse lapse histories were assigned to each cluster suggests Multivariate withdrawal model that symptom shape does not simply parrot postcessation smoking experiences. Elevation 0.99 0.26 13.45 <.001 2.67 per SD Affect group 1.45 0.46 9.86 .002 4.24 Urge group 1.37 0.62 4.94 .026 3.93 Symptom Slopes and Relapse

Contribution of nonwithdrawal variables Table 2 presents the results of the foregoing logistic regression Elevation 0.99 0.27 13.21 <.001 2.70 per SD models, with simple slope measures substituted for cluster distinc- Affect group 1.34 0.47 8.13 .004 3.85 tions. In each model, slope measures remained significantly asso- Urge group 1.39 0.63 4.87 .027 4.03 ciated with relapse, suggesting that symptom-relapse associations Plasma nicotine 0.53 0.25 4.51 .033 1.70 per SD did not depend on clustering per se or on a particular sample split. Lapse-adjusted model Some differences in other model coefficients emerged when slopes were used. Most notably, the odds ratio associated with the ele- Lapse status 4.51 0.71 40.60 <.001 91.00 vation variable was increased in all models, and plasma nicotine Elevation 0.99 0.37 7.31 .006 2.69 per SD Affect group 1.84 0.68 7.44 .006 6.34 was no longer a significant predictor of relapse. As with the cluster classifications, Hosmer-Lemeshow goodness-of-fit statistics indi- Note. The odds ratio (OR) for the elevation and plasma nicotine measures cated that all three models fit the data adequately: multivariate reflects the OR for a 1-SD increase on each of these continuous measures. withdrawal model, ~(8, N = 159) = 10.74, p = .22; nonwith- Reference groups for categorical variables were the prototypic groups for the symptom measures and nonlapsers for lapse status. For all measures, drawal variables model, ~(8, N = 157) --- 10.19, p --- .25; lapse- higher values of the OR indicate a higher probability of being counted as adjusted model, ~(8, N = 159) = 9.82, p = .28. relapsed at the Day 30 assessment. Parameter estimates reported for all When elevation and affect slopes were forced into a logistic three analyses reflect results with all listed variables simultaneously in- model of relapse, the prototypic-atypical cluster distinction im- cluded in the model. proved the model when they were entered at a subsequent step, model )(2(1, N = 159) = 3.68, p = .05. This suggests that cluster distinctions may capture features of withdrawal profiles not re- A logistic regression analysis was performed to determine flected in slopes and that these features are related to relapse. The whether any withdrawal variables could improve on a prediction urge cluster distinction did not significantly improve a model model once intrastudy smoking lapses were included as a covari- incorporating slope and elevation information, model 3(2(1, N = ate. As in the multivariate model, all withdrawal parameters except 159) = 1.88, ns. This makes sense when considering the two sleep/energy clusters were considered as candidate predictors. cluster profile shapes (see Figure 2); urge cluster group profiles Lapse status was entered along with the remaining withdrawal differed chiefly in slope per se across time. parameters to form a saturated model, and withdrawal variables were deleted as necessary. Urge classification, 9~2(1, N = External Correlates 159) = 1.85, p > .05, was deleted without a significant decrement in model fit. Affect classification and symptom elevation could not Elevation was related to FTQ scores (r =. 17, p < .05), negative be deleted without worsening overall model fit, so they were PANAS scores (r = .40,p < .01), CESD scores (r = .35,p < .01), retained. The bottom portion of Table 1 displays the results of the and PS scores (r =. 16, p < .05). No significant differences were final model. The Hosmer-Lemeshow fit statistic for this final found between prototypic and atypical cluster members in either lapse-adjusted model suggested that the model fit the data well, the affect or urge analyses on the smoker characteristic measures ~(8, N = 159) = 7.48, p = .49. (plasma nicotine, plasma cotinine, FTQ, age of smoking initiation, In a final set of assessments, clusters were divided into sub- regular smoking rate, number of quit attempts, baseline PANAS, groups according to the latency at which the first lapse was recorded. This was done only for affect and urge cluster solutions, because these were the only symptom domains associated with 4 Subgroup tallies for each cluster (listed as n continuously abstinent, n relapse. In each cluster, at least five of the six possible lapse reporting first lapse on Day 1, n reporting on Day 2, n reporting on Day 7, histories were represented; some continuously abstinent smokers n reporting on Day 14, and n reporting on Day 30, respectively) were as were found in each cluster.'* Moreover, intracluster group profiles follows: Affect I, 13, 26, 4, 6, 1, 0; Affect If, 5, 25, 2, 2, 0, 2; Affect III, 6, showed fairly tight correspondence across time. Coefficient alphas 22, 3, 8, 2, 1; Affect IV, 5, 19, 2, 3, 1, 1; Urge I, 26, 69, 5, 13, 2, 3; Urge I/, 3, 23, 6, 6, 2, 1. Clearly, lapse reports were not evenly distributed across across groups within each cluster were as follows: Affect I, .89; the postcessation period within clusters, and clusters differed from one Affect II, .95; Affect III, .87; Affect IV, .93; Urge I, .98; Urge II, another in subgroup composition. The intracluster differences in subgroup .78. Figure 3 illustrates this correspondence, depicting the intra- size, at least in part, reflect the lapse history of the sample, with early lapse cluster withdrawal profiles as a function of latency to first reported reports being quite common. Cross-cluster differences, especially in the lapse for the affect clusters. Two interpretations stem from these number of continuously abstinent persons, are congruent with the results of descriptive findings. First, the tight coherence among within- the logistic regression models. 82 PIASECKI ET AL.

7 CLUSTER I CLUSTER II A 6 k .~ td5 84

I I I I I I I I [ I QD 2 7 14 30 QD 2 7 14 30 DAY DAY

I CLUSTER III ] 7 CLUSTER IV ] 6 / ~. 5 o ...... •~. ../ 8 ¢t3 . :' cJ, ,- 4 ~.4 _ %,,-'" .~.f.' ,

2 2

L I I I I I I L I I QD 2 7 14 30 QD 2 7 14 30 DAY DAY

First Lapse Reported: Continuously Abstinent ...... Day 7 Visit Day 1 Visit ...... Day 14 Visit Day 2 Visit ...... Day 30 Visit

Figure 3. Affect score profiles for subgroups formed on the basis of latency to first reported cigarette in each affect cluster. Profiles are depicted in factor score units. QD = quit day. (Subgroup ns are given in Footnote 4.)

CESD, and PS). None of these measures was significantly corre- Examination of temporal profiles in three distinct symptom lated with affect or urge slopes. No connections between preces- domains revealed that a substantial proportion of the sample sation variables and either cluster distinctions or slopes were found reported atypical withdrawal patterns for each symptom class. in models with elevation statistically controlled. After controlling for overall symptom severity (elevation), the prototypic-atypical distinction for the Affect and Urge subscales Discussion was found to be associated with smoking status at follow-up. After lapse status was statistically controlled, affect classification still One major aim of this research was to evaluate the robustness of significantly improved the logistic model of relapse. Temporal two earlier findings--the existence of marked heterogeneity in variability in Affect factor scores was more robustly related to postcessation symptom patterns, and the connection between tem- clinical outcomes than was variability in other withdrawal subdo- poral symptom parameters and relapse (Piasecki et al., 1998). A mains. This is consistent with our speculation that affective pro- conceptual replication of these findings was deemed especially cesses play a central motivational role in the return to smoking. important given the historical inconsistency of withdrawal mea- Withdrawal variables considered alone constituted an adequate sures (Patten & Martin, 1996a). Because we sought a relatively "risky" test of these phenomena, we examined data from a study fitting model of relapse likelihood, but measures of baseline that was methodologically quite different from our previous stud- plasma nicotine and cotinine yielded an orthogonal improvement ies. This sample comprised unaided quitters, used a different of this model. This and other findings suggest that temporal withdrawal measure, and assessed withdrawal relatively infre- variance in symptom measures does not simply mediate the influ- quently during a shorter postcessation window. Despite these ence of nicotine dependence. Finally, connections between symp- differences, we obtained evidence corroborating earlier findings. tom patterns and relapse were robust, regardless of whether they SMOKING WITHDRAWAL 83

Table 2 affect self-reports may exert somewhat independent influences on Logistic Regression Results Assessing Relations Between relapse. It is possible, for instance, that some smokers may be urge Withdrawal Variables and Relapse, Substituting Simple reactive whereas others are affect reactive. This conjecture must Slope Measures for Cluster Distinctions remain speculative in lieu of further research. Descriptive research has shown that urges are often higher Predictor [3 SE (/3) Wald p OR during continued smoking than during an abstinence attempt (e.g., Multivariate withdrawal model Hughes, 1992). Hence, the pattern of increasing urge scores seen in the atypical urge group in this sample may simply reflect the Elevation 1.21 0.29 17.82 <.001 3.37 per SD reestablishment of typical smoking behavior across the study pe- Affect slope 0.89 0.26 11.64 <.001 2.44 per SD Urge slope 0.70 0.27 7.08 .008 2.02 per SD riod. This might explain why, when the study controlled for lapses, the relation between urge cluster membership and relapse was Contribution of nonwithdrawal variables attenuated. If smoking is associated with elevated urges, a measure Elevation 1.18 0.29 16.62 <.001 3.27 per SD of lapse occurrence would be statistically redundant with urge Affect slope 0.80 0.26 9.34 .002 2.22 per SD growth, causing urge classification to drop out of the prediction Urge slope 0.62 0.27 5.18 .023 1.85 per SD model. Of course, there is independent evidence that urge ratings Plasma nicotine 0.40 0.24 2.81 .093 1.49 per SD acutely foreshadow first lapses (e.g., Shiffman, Paty, Gnys, Kassel, Lapse-adjusted model & Hickcox, 1996; Shiffman et al., 1997). A great deal of evidence links affect to smoking motivation and Lapse status 4.53 0.72 39.82 <.001 92.79 demonstrates that affect may be modified by smoking (Brandon, Elevation 1.21 0.40 9.03 .003 3.35 per SD Affect slope 1.04 0.38 7.37 .006 2.82 per SD 1994). Nonetheless, it seems likely that patterns of affect ratings after cessation reflect more than simple smoking status. For in- Note. The odds ratio (OR) for all predictors except lapse status reflects stance, affect patterns may reflect the cumulation of stressors, the the OR for a 1-SD increase on each of these continuous measures. The mood-related consequences of progressive sleep deprivation, the reference group for lapse status is nonlapsers. For all measures, higher loss of reinforcement, the emergence of a depressive episode, and values of the OR indicate a higher probability of being counted as relapsed at the Day 30 assessment. Parameter estimates reported for all three so on. Postcessation lapses may acutely alleviate exacerbations in analyses reflect results with all listed variables simultaneously included in negative affect, but, in many cases, smoking would not be ex- the model. pected to remove the source of mounting dysphoria. Ex-smokers seeking more consistent relief from negative mood may turn to cigarettes more and more frequently, eventuating in relapse. This were represented as cluster-based categorical variables or simple account is likewise consistent with the finding that affect classifi- slope measures. This suggests that the findings are not an artifact cation contributed to a logistic model of relapse even after early associated with any particular data reduction technique. Thus, this lapses were controlled. research replicates our earlier work that showed that mean symp- Another major aim of this research was to investigate preces- tom elevation and the trajectory of symptoms are strongly associ- sation variables that might foretell which individuals are at risk for ated with smokers' eventual relapse fates. These findings differ atypical symptom dynamics. Although an array of precessation markedly from conclusions of many existing studies (Hughes et measures was available to correlate with cluster membership and al., 1990; Patten & Martin, 1996a). symptom slopes (Shadel et al., 1998), we restricted our precessa- Earlier, we noted that applying unique analytic techniques to tion analyses to dependence and affect measures for theoretical accepted measures may throw new light on the meaning of those reasons. These analyses revealed no significant relations between measures. What is the core of the meaningful information in these measures and either cluster membership or symptom slopes. withdrawal measures? We constructed symptom subscales that Symptom elevation was related to baseline PTQ scores and a were essentially orthogonal to one another in an attempt to isolate variety of affect measures. This finding corresponds with results of the core, clinically relevant sources of variance in withdrawal a recent study by Gilbert and colleagues (1998) that show robust scales. Plotting these scales against time in the aggregate showed connections between trait measures of affect and the broad-span that only urge and affect ratings conformed to conventional ex- severity of abstinence symptomatology. The dearth of connections pectations regarding the time course of withdrawal symptoms. between precessation variables and various representations of Other analyses suggested that, across time, the unique variance in symptom trajectory in this sample and two previous studies (Pi- each of these measures was fairly independent. Very different asecki et al., 1998), coupled with the prediction of severity from profile shapes were obtained for the two symptom classes in baseline measures, suggests that postcessation events and stimuli cluster analyses. Of course, this may result in part from metric may account for the dynamic features of abstinence symptomatol- properties of the computed scales or from the imprecision inherent ogy. This predictive dissociation and the unique contributions of in cluster analysis. However, the finding that cluster groupings elevation and trajectory to logistic models of relapse confirm the based on both urge and affect were associated with relapse when independence of these two facets of withdrawal. entered simultaneously into logistic regression models suggests Relations between smoking and symptom measures in this re- that the two measures may reflect different processes. This con- search were complex, and these connections are in need of further clusion is supported by the fact that membership in the atypical research. Symptom measures were robustly related to smoking groups differed substantially across the two measures. Such find- status at follow-up, suggesting a strong connection between the ings suggest that factors that give rise to trajectories of urge and two variables. Lapses might directly cause both relapse and atyp- 84 PIASECKI ET AL. ical withdrawal patterns. A strong version of a lapse-based account shape may be too complex and rich to be captured by a simple might ascribe little or no causal role for withdrawal patterns in change index. relapse. Ancillary analyses aimed at characterizing the relations Limitations of our study must be considered in evaluating this between lapses, 30-day outcomes, and symptom patterns challenge research. For instance, the follow-up window in this study was such an account. Smokers classified in typical versus atypical short, and the point-prevalence window for the abstinence criterion groups did not differ in the pattern or rate of smoking after overlapped with the period in which withdrawal was being as- cessation. Moreover, including lapse status in logistic models as a sessed. These aspects of the study prevent us from definitively covariate did not eliminate the relation between affect symptom assessing the temporal precedence of symptom change in the patterning and smoking status at follow-up, suggesting that with- return to smoking. Other limitations of the study include the fairly drawal dynamics contribute nonredundant information regarding small sample size, the fact that dally symptom data were not relapse risk. Finally, continuously abstinent smokers and smokers collected, the exclusive use of self-report measures, and the brief with diverse latencies to the first lapse were represented in each postcessation study period (i.e., only 30 days). The small sample cluster. This demonstrates that symptom pattern differences do not size means that some of our findings may not generalize well to depend upon particular lapse histories. self-quitters. The sample comprised self-quitters, which we viewed One could legitimately question whether symptom reports gath- as an important distinction from our past studies of nicotine patch ered long after cessation truly reflect withdrawal. One response users. Nonetheless, because they were willing to commit to time- would be to equate concretely the construct with the measures and consuming data collection procedures, the present sample may not claim that the analyses must capture withdrawal if they are based be highly representative of self-quitters in general. Moreover, the on widely used and validated instruments. This argument is, of lack of daily data and the brief follow-up period certainly affected course, simplistic. A clear .conceptual definition of withdrawal is the shapes of the postcessation profiles obtained. This also means required to address this question. For instance, if lapses trigger that the profile shapes cannot be easily compared with those larger symptomatic increases in abstinent versus continuing smok- generated in our previous research (Piasecki et al., 1998). In ers, would such symptomatic reactions to drug use constitute addition, some of the smokers labelled as abstinent at 30-days withdrawal responses? Moreover, if smoking is used to suppress postcessation certainly went on to relapse at a later time (Gilpin, Pierce, & Farkas, 1997). The lack of objective (i.e., non-self- the emergence of affective disorder, is the affective disorder a report) data means that we know virtually nothing about the withdrawal symptom complex? At present, the field's data- physiological and behavioral changes that smokers are undergoing gathering and analytic capabilities may have outstripped the com- late in the postcessation period. For instance, we do not know plexity and clarity of its definitions. whether smokers are actually experiencing greater neural activity Symptom trajectory was related to follow-up smoking status in in systems that process negative affect; thus, the exacerbation in this research, regardless of whether it was represented by categor- rated withdrawal symptoms reported by some participants may ical cluster assignments or continuous slope measures. This natu- merely reflect changes in their perception or scaling of affective rally raises questions about the value of cluster analysis and other signals. Reactive effects associated with repeated self-reports of categorical data reduction methods in withdrawal heterogeneity mood states may also have contributed to some observed symp- research. Can cluster analysis and similar techniques contribute tomatic change (Gilbert et al., 1998). Future research should be anything to the study of withdrawal that slopes cannot? The best designed to assess these concerns, incorporating precessation statistical representation of trajectory variance will need to be run-in assessments to minimize subsequent reactivity and includ- scrutinized in future research. We suspect that continuous mea- ing periodic objective assessments after the quit date. sures (simple slopes, correlations between individual profiles, and theoretical function forms, etc.) will be found to be useful tools for References assessing relations among symptom trajectory, precessation variables, postcessation events, and clinical outcomes. Nonetheless, American Psychiatric Association. (1994). Diagnostic and statistical man- cluster analyses may continue to be valuable at initial stages of the ual ofrnental disorders (4th ed.). Washington, DC: Author. data analytic process, providing useful descriptive information not Baker, T. B., Morse, E., & Sherman, J. E. (1987). The motivation to use captured by other representations of trajectory. Clustering isolates drugs: A psychobiological analysis of urges. In C. Rivers (Ed.), Ne- braska symposium on motivation (Vol. 34, pp. 257-323). Lincoln: a family of common profiles present in a sample, and these may University of Nebraska Press. have many useful diagnostic applications. For instance, modal Brandon, T. H. (1994). Negative affect as motivation to smoke. Current symptom profiles can provide a sense of the frequency with which Directions in Psychological Science, 3, 33-37. particular patterns are observed, and, indeed, certain profiles might Brandon, T. H., Tiffany, S. T., Obremski, K. M., & Baker, T. B. (1990). prove to have taxonic significance. Moreover, examining these Postcessation cigarette use: The process of relapse. Addictive Behav- profiles may help researchers locate particular time periods in iors, 15, 105-114. which symptoms worsen, permitting hypothesis generation and the Buck, R. (1993). What is this thing called subjective experience? Reflec- construction of sensitive continuous indicators of trajectory. In the tions on the neuropsychology of qualia. Neuropsychology, 7, 490-499. Cattell, R. B. (1966). The data box: Its ordering of total resources in terms present research, affect cluster distinctions were found to contrib- of possible relational systems. In R. B. Cattell (Ed.), Handbook of ute information to a model of relapse that controlled for symptom multivariate experimental psychology (pp. 67-128). Chicago: Rand- elevation and affect slope. We attribute this finding to the sensi- McNally. tivity of clustering to transient but potentially important vicissi- Cohen, S., Kamarck, T., & Mermelstein, R. (1983). A global measure of tudes in withdrawN. This is consistent with the fact that profile perceived stress. Journal of Health and Social Behavior, 24, 385-396. SMOKING WITHDRAWAL 85

Cronbach, L. J., & Gleser, G. C. (1953). Assessing similarity between Marlatt, G. A. (1985). Relapse prevention: Theoretical rationale and over- profiles. Psychological Bulletin, 50, 456-473. view of the model. In G. A. Marlatt & J. R. Gordon (Eds.), Relapse Epping-Jordan, M. P., Watkins, S. S., Koob, G. F., & Markou, A. (1998). prevention: Maintenance strategies in the treatment of addictive behav- Dramatic decreases in brain reward function during nicotine withdrawal. iors (pp. 3-70). New York: Guilford Press. Nature, 393, 76-79. Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, Sir Fagerstrtm, K. O. (1978). Measuring degree of physical dependence to Ronald, and the slow progress of soft psychology. Journal of Consulting tobacco smoking with reference to individualization of treatment. Ad- and Clinical Psychology, 46, 806-834. dictive Behaviors, 3, 235-241. Mortal, A. R., Iguchi, M. Y., Belding, M. A., & Lamb, R. J. (1997). Fiore, M. C., Novotny, T. E., Pierce, J. P., Giovino, G. A., Hatziendreu, Natural classes of treatment response. Journal of Consulting and Clin- E. J., Newcomb, P. A., Surawicz, T. S., & Davis, R. M. (1990). Methods ical Psychology, 65, 673-685. used to quit smoking in the United States: Do cessation programs help? Overall, J. E., & Magee, K. N. (1992). Replication as a role for determining Journal of the American Medical Association, 263, 2760-2765. the number of clusters in hierarchical cluster analysis. Applied Psycha- Gilbert, D. G., McClemon, F. J., Rabinovich, N. E., Plath, L. C., Jensen, logical Measurement, 16, 119-128. R. A., & Meliska, C. J. (1998). Effects of smoking abstinence on mood Patten, C. A., & Martin, J. E. (1996a). Does nicotine withdrawal affect and craving in men: Influences of negative-affect-related personality smoking cessation? Clinical and theoretical issues. Annals of Behavioral traits, habitual nicotine intake, and repeated measurements. Personality Medicine, 18, 190-200. and Individual Differences, 25, 399-423. Patten, C. A., & Martin, J. E. (1996b). Measuring tobacco withdrawal: A Gilpin, E. A., Pierce, J. P., & Farkas, A. J. (1997). Duration of smoking review of self-report questionnaires. Journal of Substance Abuse, 8, abstinence and success in quitting. Journal of the National Cancer 93-113. Institute, 89, 572-576. Piasecki, T. M., Fiore, M. C., & Baker, T. B. (1998). Profiles in discour- Gourlay, S. G., Forbes, A., Marriner, T., Pethica, D., & McNeil, J. J. agement: Two studies of variability in the time course of smoking (1994). Prospective study of factors predicting outcome of transdermal withdrawal symptoms. Journal of Abnormal Psychology, 107, 238-251. nicotine treatment in smoking cessation. British Medical Journal 309, Piasecki, T. M., Kenford, S. L., Smith, S. S., Fiore, M. C., & Baker, T. B. 842- 846. (1997). Listening to nicotine: Negative affect and the smoking with- Heyman, G. M. (1996). Resolving the contradictions of addiction. Behav- drawal conundrum. Psychological Science, 8, 184-189. ioral and Brain Sciences, 19, 561-574. Prochaska, J. O., Velicer, W. F., DiClimente, C. C., Guadagnoli, E., & Hosmer, D. W., & Lemeshow, S. (1989). Applied logistic regression. New Rossi, J. S. (1991). Patterns of change: Dynamic typology applied to York: Wiley. smoking cessation. Multivariate Behavioral Research, 26, 83-107. Hughes, J. R. (1992). Tobacco withdrawal in self-quitters. Journal of Radloff, L. S. (1977). The CES-D scale: A self-report depression scale for Consulting arm Clinical Psychology, 60, 689-697. research in the general population. Applied Psychological Measure- Hughes, J. R., & Hatsukami, D. K. (1986). Signs and symptoms of tobacco ment, 1, 385-401. withdrawal. Archives of General Psychiatry, 43, 289-294. Robinson, T. E., & Berridge, K. C. (1993). The neural basis of drug Hughes, J. R., Higgins, S. T., & Hatsukami, D. K. (1990). Effects of craving: An incentive-sensitization theory of addiction. Brain Research abstinence from tobacco: A critical review. In L. T. Kozlowski, H. Reviews, 18, 247-291. Armis, H. D. Cappell, F. Glaser, M. Goodstasdt, Y. Israel, H. Kalant, Schweizer, K. (1992). A correlation-based decision-rule for determining E. M. Sellers, & J. Vingillis (Eds.), Research advances in alcohol and the number of clusters and its efficiency in uni- and multi-level data. drug problems (Vol. 10, pp. 317-398). New York: Plenum Press. Multivariate Behavioral Research, 27, 77-94. Isbell, H., & White, W. M. (1953). Clinical characteristics of addictions. Seevers, M. H. (1962). Medical perspectives on habituation and addiction. American Journal of Medicine, 17, 558-565. Journal of the American Medical Association, 181, 92-98. Jaffe, J. H. (1980). Drug addiction and drug abuse. In A. G. Gilman, L. S. Serlin, R. C., & Lapsley, D. K. (1985). Rationality in psychological Goodman, & A. Gilman (Eds.), Goodman & Gilman "s: The pharmaco- research: The good-enough principle. American Psychologist, 40, 73- logical basis of therapeutics (6th ed.). New York: Macmillan. 83. Jorenby, D. E., Hatsukami, D. K., Smith, S. S., Fiore, M. C., Allen, S., Shadel, W. G., Niaura, R., Abrams, D. B., Goldstein, M. G., Rohsenow, Jensen, J., & Baker, T. B. (1996). Characterization of tobacco with- D. J., Sirota, A. D., & Monti, P. M. (1998). Scripted imagery manipu- drawal symptoms: Transdermal nicotine reduces hunger and weight lations and smoking cue reactivity in a clinical sample of self-quitters. gain. Psychopharmacology, 128, 130-138. Experimental and Clinical Psychapharmacology, 6, 179-186. Jorenby, D. E., Smith, S. S., Fiore, M. C., Hurt, R. D., Offord, K. P., Shiffman, S., Enberg, J. B., Paty, J. A., Perz, W. G., Gnys, M., Kassel, Croghan, I. T., Hays, J. T., Lewis, S. F., & Baker, T. B. (1995). Varying J. D., & Hickcox, M. (1997). A day at a time: Predicting smoking lapse nicotine patch dose and type of smoking counseling. Journal of the from daily urge. Journal of Abnormal Psychology, 106, 104-116. American Medical Association, 274, 1347-1352. Shiffman, S. M., & Jarvik, M. E. (1976). Smoking withdrawal symptoms Keenan, R. M., Hatsukami, D. K., Pentel, P. R., Thompson, T. N., & Grillo, in two weeks of abstinence. Psychopharmacology, 50, 35-39. M. A. (1994). Pharmacodynamic effects of cotinine in abstinent cigarette Shiffman, S., Paty, J. A., Gnys, M., Kassel, J. A., & Hickcox, M. (1996). smokers. Clinical Pharmacology and Therapeutics, 55, 581-590. First lapses to smoking: Within-subjects analysis of real-time reports. Kenford, S. L., Fiore, M. C., Jorenby, D. E., Smith, S. S., Wetter, D., & Journal of Consulting and Clinical Psychology, 64, 366-379. Baker, T. B. (1994). Predicting smoking cessation: Who will quit with Skinner, H. A. (1981). Toward the integration of classification theory and and without the nicotine patch. Journal of the American Medical Asso- methods. Journal of Abnormal Psychology, 90, 68-87. ciation, 271, 589-594. Skinner, H. A. (1990). Validation of the dependence syndrome: Have we Krolak-Schwerdt, S., & Eckes, T. (1992). A graph-theoretic criterion for crossed the half-life of this concept? In G. Edwards & M. Lader (Eds.), determining the number of clusters in a data set. Multivariate Behavioral The nature of drug dependence (pp. 41-73). Oxford, England: Oxford Research, 27, 541-565. University Press. Lichtenstein, E., & Hollis, J. F. (1992). Patient referral to a smoking Swan, G. E., Ward, M. M., & Jack, L. M. (1996). Abstinence effects as cessation program: Who follows through? Journal of Family Prac- predictors of 28-day relapse in smokers. Addictive Behaviors, 21, 481- tice, 34, 739-744. 490. 86 PIASECKI ET AL.

Transderrnal Nicotine Study Group. (1991). Transdermal nicotine for Baker, T. B. (1999). Development and validation of the Wisconsin smoking cessation: Six month results from two multicenter controlled Smoking Withdrawal Scale. Experimental and Clinical Psychopharma- clinical trials. Journal of the American Medical Association, 266, 3133- cology, 7, 354-361. 3138. Wetter, D. W., Fiore, M. C., Baker, T. B., & Young, T. B. (1995). Vuchinich, R. E., & Tucker, J. A. (1988). Contributions from behavioral Tobacco withdrawal and nicotine replacement influence objective theories of choice to an analysis of alcohol abuse. Journal of Abnormal measures of sleep. Journal of Consulting and Clinical Psychol- Psychology, 97, 181-195. ogy, 63, 658-667. Ward, J. (1963). Hierarchical grouping to optimize an objective function. Wikler, A. (1973). Dynamics of drug dependence: Implications of a con- Journal of the American Statistical Association, 58, 236-244. ditioning theory for research and treatment. Archives of General Psy- Watson, D., Clark, L. A., & TeUegen, A. (1988). Development and vali- chiatry, 28, 611-616. dation of brief measures of positive and negative affect: The PANAS Wood, J. M., Tataryn, D. J., & Gorsuch, R. L. (1996). Effects of under- and scales. Journal of Personality and Social Psychology, 54, 1063-1070. overextraction on principal axis factor analysis with varimax rotation. Welsch, S. K., Smith, S. S., Jorenby, D. E., Wetter, D. W., Fiore, M. C., & Psychological Methods, 1, 354-365.

Appendix

Items Constituting Each Subscale and Their Loadings on Those Factors

Item Loading

Factor 1: Affect Irritability .826 Anxiety .787 Impatience .779 Restlessness .752 Difficulty concentrating .684 Sadness .603 Appetite increase .476

Factor 2: Urge Urge to smoke today .760 Miss a cigarette .752 Wanting/craving a cigarette .749 If could smoke freely, want cigarette now .729 Thinking of cigarettes more than usual .715 If permitted to smoke, refuse cigarette now a .644

Factor 3: Sleep/Energy Disrupted sleep .663 Trouble falling asleep .654 Drowsiness .634 Dreaming more than usual .623 Trouble waking .605 Decreased heart rate .598 Muscle tension .516

a Reverse scored in all analyses. Received September 17, 1997 Revision received April 26, 1999 Accepted April 26, 1999 •